Changelog #

12.0.0 #

Require the 12.0.0 package suite; no additional public API changes.

11.0.0 #

Update dependency constraints for the 11.0.0 suite.

10.0.0 #

Update dependency constraints for the 10.0.0 suite; no public API changes in this package.

9.0.0 #

Add an optional detection pixel-area cap (PpOcrPipeline.detectionPixelLimit, detectionResize(pixelLimit:)) for memory-bound backends. This fixes browser OCR on iOS.
Align dependency constraints with the dart-pdf 9.0.0 suite.

8.1.0 #

Align dependency constraints with the dart-pdf 8.1.0 suite.

8.0.0 #

Align dependency constraints with the dart-pdf 8.0.0 suite.

7.0.0 #

Much better accuracy on small print and technical drawings, and ~4x faster. On a synthetic A3 signalling sheet (buildOcrDrawingSheet in pdf_test_fixtures) exact label reads go from 18% to 95% (CER 69% -> 1.4%).
- Detection no longer squeezes the page to 960 px: detectionSideLimit defaults to 4000 (PaddleOCR's PP-OCRv5 cap). At 960 a drawing's ~5pt labels fused with the symbols beside them and read as garbage.
- DB unclip is DB's own distance offset (area * ratio / perimeter, ratio 1.5) instead of scaling the box 1.6x about its centre; box score threshold 0.6 (PaddleOCR's).
- Recognition keeps each line's aspect, padding to at least 320 px (recognitionInput(minWidth:); the record gains paddedWidth), instead of squashing every line into 512 px. OnnxOcrModelRunner gains recognitionMinWidth; recognitionMaxWidth is now only a safety cap (default 3200).
- Recognized lines go through cleanRecognizedText: drawing symbols (Geometric Shapes, ● ▲ ▼) are removed and symbol-only lines dropped (OnnxOcrModelRunner(cleanText: false) keeps the raw output).
- New OnDeviceOcrEngine.pixelRatioFor(page): 216 dpi, capped at 4000 px a side - pass it to applyOcr instead of pixelRatio: 2.
- Model outputs are read straight from ONNX Runtime's buffer instead of through OrtValueTensor.value's nested lists, which cost ~3x the inference itself. Adds a direct ffi dependency.
New web-safe library package:pdf_ocr_ondevice/pp_ocr.dart: the pipeline is now PpOcrPipeline (plain Dart) over a PpOcrInference backend - OnnxOcrModelRunner is that pipeline over native ONNX Runtime, and a browser can supply onnxruntime-web (the DartPDF app now does, in place of Florence-2). OcrRunnerEngine is the backend-agnostic PdfOcrEngine (OnDeviceOcrEngine extends it; pixelRatioFor lives there too).
Recognition runs in aspect-sorted batches of 6 lines (recognitionBatchSize, PaddleOCR's default), each padded to its own widest line: identical readings, page time halved again (37 s -> 19 s on the benchmark sheet).
Raise the Flutter floor to flutter: '>=3.47.0', matching dart_pdf_editor 7.0.0 (material_ui 1.4's floor).

6.0.0 #

Declare the real Flutter floor: flutter: '>=3.44.0' (was >=3.24.0), matching dart_pdf_editor. 5.0.0 already needed Flutter 3.44 (ReorderableListView.onReorderItem), so this strands nobody.

5.1.1 #

Version bump to track the 5.1.1 suite.

5.1.0 #

Version bump to track the 5.1.0 suite.

5.0.0 #

Version bump to track the 5.0.0 suite.

4.5.0 #

Align dependency constraints with the dart-pdf 4.5.0 package suite.

4.4.0 #

Align dependency constraints with the dart-pdf 4.4.0 package suite.

4.3.0 #

Align dependency constraints with the dart-pdf 4.3.0 package suite.

4.2.0 #

Lockstep minor release aligned with dart_pdf_editor 4.2.0. No public on-device OCR API changes since 4.1.0.

4.1.0 #

Lockstep minor release aligned with dart_pdf_editor 4.1.0. No public on-device OCR API changes since 4.0.0.

4.0.0 #

Lockstep major release to align with dart_pdf_editor 4.0.0. No public on-device OCR API changes since 3.8.0.

3.8.0 #

Lockstep minor release to align with dart_pdf_editor 3.8.0. No public on-device OCR API changes since 3.7.0.

3.7.0 #

Lockstep minor release to align with dart_pdf_editor 3.7.0. No public on-device OCR API changes since 3.6.0.

3.6.0 #

Lockstep minor release to align with dart_pdf_editor 3.6.0. No public on-device OCR API changes since 3.5.1.

3.5.1 #

Lockstep patch release to align with dart_pdf_editor 3.5.1. No public on-device OCR API changes since 3.5.0.

3.5.0 #

Lockstep minor release to align with dart_pdf_editor 3.5.0. No public on-device OCR API changes since 3.4.0.

3.4.0 #

Lockstep minor release to align with dart_pdf_editor 3.4.0. No public on-device OCR API changes since 3.3.1.

3.3.1 #

Lockstep patch release to align with dart_pdf_editor 3.3.1. No public on-device OCR API changes since 3.3.0.

3.3.0 #

Lockstep minor release to align with dart_pdf_editor 3.3.0. No public on-device OCR API changes since 3.2.0.

3.2.0 #

Lockstep minor release to align with dart_pdf_editor 3.2.0. No public on-device OCR API changes since 3.1.1.

3.1.1 #

Lockstep patch release to align with dart_pdf_editor 3.1.1. No public on-device OCR API changes since 3.1.0.

3.1.0 #

Lockstep release to align with dart_pdf_editor 3.1.0. No public on-device OCR API changes since 3.0.0.

3.0.0 #

Lockstep major release to align with dart_pdf_editor 3.0.0. No public on-device OCR API changes since 2.1.0.

2.1.0 #

Version bump to align with dart_pdf_editor 2.1.0. No public on-device OCR API changes since 2.0.0.

2.0.0 #

Version bump to align with dart_pdf_editor 2.0.0. No public on-device OCR API changes since 1.4.7.

1.4.7 #

Version bump to align with dart_pdf_editor 1.4.7. No public on-device OCR API changes since 1.4.6.

1.4.6 #

Version bump to align with dart_pdf_editor 1.4.6. No public on-device OCR API changes since 1.4.5.

1.4.5 #

Version bump to align with dart_pdf_editor 1.4.5. No public on-device OCR API changes since 1.4.4.

1.4.4 #

Version bump to align with dart_pdf_editor 1.4.4. No public on-device OCR API changes since 1.4.3.

1.4.3 #

Version bump to align with dart_pdf_editor 1.4.3. No public on-device OCR API changes since 1.4.2.

1.4.2 #

Version bump to align with dart_pdf_editor 1.4.2. No public on-device OCR API changes since 1.4.1.

1.4.1 #

Run downloaded ONNX OCR models on a long-lived worker isolate by default. Page RGBA buffers transfer without a structured-message copy, and model loading, preprocessing, inference, and post-processing no longer block the UI isolate.

1.4.0 #

Version bump to align with dart_pdf_editor 1.4.0. Runtime maintenance keeps the on-device OCR package compatible with the 1.4.0 editor/app release.

1.3.2 #

Version bump to align with dart_pdf_editor 1.3.2. No API changes since 1.3.1.

1.3.1 #

Version bump to align with dart_pdf_editor 1.3.1. No API changes since 1.2.3.

1.2.3 #

Added PdfOcrDownloadCancelToken so hosts can wire a Cancel button to in-flight model downloads; cancellation removes partial files and throws PdfOcrModelDownloadCanceled.
Expanded the README download example to display file/overall progress and show where to call cancel().
Fix on-device OCR failing on Windows with a garbled "Load model from … File doesn't exist" error. The ONNX Runtime session is now created from the model bytes (OrtSession.fromBuffer) instead of a file path: on Windows the binding passed the path as a narrow UTF-8 string where ONNX Runtime expects a wide wchar_t*, mangling every path - even pure-ASCII ones - into CJK mojibake. This supersedes the 1.2.2 ASCII path-staging workaround, which could not help because the corruption happened regardless of the path's contents.

1.2.2 #

Fix on-device OCR model path staging on Windows so the downloaded PP-OCR ONNX model resolves correctly.

1.2.1 #

Add a package example and shorten the pubspec description for pub.dev scoring.

1.2.0 #

Downloadable on-device OCR package for the DartPDF app, with model-manager integration and native-platform OCR engine wiring.
Version bump to align with dart_pdf_editor 1.2.0.

0.1.0 #

Initial release. On-device, downloadable OCR for dart_pdf_editor.
PdfOcrModelManager downloads, caches (under the app-support directory), integrity-checks (SHA-256), and removes OCR model bundles, reporting progress as bytes arrive. Native platforms only (isSupported is false on the web).
OnDeviceOcrEngine implements PdfOcrEngine, mapping a backend's pixel-space text lines into PDF user space. PdfEditor.applyOcr writes an invisible, selectable layer with no per-page network call.
OnnxOcrModelRunner runs a PP-OCR detect+recognize pipeline on ONNX Runtime (det resize/normalize, DB box extraction, CRNN/CTC decode), all of the pre/post-processing in pure, unit-tested Dart.
PdfOcrModels.ppOcrV5Mobile describes the recommended lightweight model; point its file URLs at a bundle you host (see the README).

pdf_ocr_ondevice 12.0.0 pdf_ocr_ondevice: ^12.0.0 copied to clipboard

Metadata